Attested analytics · regulated finance

Every number, proven — not promised.

Ask a finance question in plain English. Get the answer with the evidence it’s right: the executed SQL, an independent runtime verdict, and a signed receipt your auditors verify offline. A confidently wrong number is worse than no number — so we prove every one.

Deploy in your VPC or fully air-gappedNo data leaves your boundaryAudit-ready by default
100%
accuracy on accepted answers
the verifier guarantees it
8 / 8
finance error classes caught
intercompany, FX, grain, sign…
30.5%
model cost saved
cheap-first, escalate on doubt
$145K
error caught in the demo
before it reached a report
See it work

The whole pitch, in four beats

The third beat is the one that sells it: a wrong number, caught by name, before anyone acts on it.

“What was consolidated net revenue in 2025?
$5,293,985.00
PASS
Plain English in. The number out — with the exact SQL that produced it, run on a read-only governed warehouse.
Ask1 / 4
Why it matters

Finance already has chatbots. It doesn’t have trust.

Today’s assistants prove themselves offline — “95% accurate on a benchmark.” In a regulated report that means 5% is wrong and you don’t know which 5%, and the cost of that 5% is a restatement or an audit finding — not a shrug. So finance either distrusts the tool or pays an analyst to re-check every number, which kills self-service. Tessera removes the blocker: it proves each answer, at runtime.

The difference is structural

Not a better chatbot — the trust layer underneath it

Generic copilotText-to-SQL assistantTessera
Runs real SQL on governed datanoyesyes
Proves this answer correct, at runtimenooffline onlyyes — per answer
Catches named finance mistakesnono8 failure classes
Auditor-verifiable evidencenonosigned receipt
Fails honest (no number vs. a wrong one)nopartialWARN, never fabricates
Runs fully air-gappedrarelyrarelyyes
How it works

Ask → verify independently → attest

01

Ask

Your question resolves to a certified metric and runs as parameterized, read-only SQL — never string-built, never free-form.

02

Verify — independently

A separate engine re-derives the answer and checks it against the named ways a ledger number goes wrong. Re-running the model’s own SQL would be circular; this never does.

03

Attest

The verdict, SQL and answer are bound into an Ed25519 receipt. Anyone re-checks it offline — they trust the receipt, not the assistant.

The economics

Verified trust, at a lower bill

The verifier makes “cheap model first” safe — so you only pay for the expensive model on the questions that actually need it.

30.5%
lower model spend vs. always using the strong model, at 100% accuracy on the benchmark mix.
Your models
Pluggable tiers — run it the way your security team allows:
Ollama · air-gappedOpenAIAWS Bedrock
$0 / answer to audit
The signed receipt is the evidence trail — generated automatically, no extra reconciliation work.
Try it live

Ask the real system

This talks to the live verifier. Toggle the cost cascade and a signed receipt; open the evidence drawer and verify the receipt yourself.

See the proof for yourself.

Open-source, deploy-anywhere, and every answer carries its own evidence.